Skip to content

configuration.md

Latest commit

 

History

History
144 lines (117 loc) · 3.51 KB

configuration.md

File metadata and controls

144 lines (117 loc) · 3.51 KB

Configuration Reference

All configuration is in config.yaml. Copy from config.example.yaml to get started.

Environment Variables

Variable Required Description
GOOGLE_API_KEY Yes (if using Gemini embeddings) Google API key for gemini-embedding-001
LLM_API_KEY Yes API key for Mem0 inference and maintenance LLM
OPENAI_API_KEY Alternative to LLM_API_KEY Standard OpenAI key (used as fallback)
EMBEDDING_API_KEY Optional Separate key for embedding fallback provider
ENGRAM_CONFIG Optional Path to config.yaml (default: ../config.yaml relative to src/)
ENGRAM_HOST_CONFIG Optional Path to host config.json for provider credentials (default: ~/.mem0-gateway/config.json)
ENGRAM_DATA_DIR Optional Data directory for queues, locks, logs (default: ~/.mem0-gateway/mem0)

Agent Permissions

agents:
  main:
    read: [global, "group:*", dm, "agent:*", all, "cross:*"]
    write: [global, "group:*", dm]
    allowed_types: [preference, fact, procedure, lesson, decision, task_log, knowledge]
  worker_example:
    read: []
    write: ["agent:worker_example"]
    allowed_types: [procedure, task_log]

default_agent_policy:
  read: [global]
  write: []
  allowed_types: []

LLM Configuration (new in v2)

llm:
  model: gpt-4o-mini
  base_url: https://api.openai.com/v1
  # If using host config.json providers:
  # provider_name: your-provider

Memory Types

memory_types:
  preference: { decay: never, conflict: update_latest }
  fact: { decay: never, conflict: verify_source }
  procedure: { decay: half_life_90d, conflict: keep_latest_version }
  lesson: { decay: never, conflict: no_override }
  task_log: { decay: half_life_30d, conflict: no_dedup }
  decision: { decay: never, conflict: no_override }
  knowledge: { decay: never, conflict: merge_latest }

Maintenance

maintenance:
  daily_time: "03:00"
  weekly_day: sunday
  weekly_window_minutes: 30
  dedup_auto_threshold: 0.92
  dedup_candidate_threshold: 0.85
  llm_chain:
    - default/gpt-4o
  llm_base_url: https://api.openai.com/v1

Decay Parameters

decay:
  task_log_half_life_days: 30
  procedure_half_life_days: 90
  access_count_cap: 3.0
  archive_threshold: 0.10

Embedding

embedding:
  model: models/gemini-embedding-001
  # provider_name: your-provider  # for fallback via host config
  primary:
    provider: gemini
    model: models/gemini-embedding-001
    dimensions: 3072
  fallback:
    provider: openai
    model: gemini-embedding-001
    base_url: https://your-proxy.com/v1
    dimensions: 3072

Qdrant

qdrant:
  host: localhost     # or "qdrant" in Docker Compose
  port: 6333
  collection: memories

Safety

never_store_patterns:
  - 'sk-[A-Za-z0-9]{20,}'
  - 'AIza[0-9A-Za-z_-]{20,}'
  - 'password\s*[:=]\s*\S+'
  - 'token\s*[:=]\s*\S+'
  - 'cookie\s*[:=]\s*\S+'
  - '-----BEGIN\s+(RSA\s+)?PRIVATE'

Notification (new in v2)

notification:
  webhook_url: https://hooks.slack.com/services/xxx/yyy/zzz

Maintenance reports and alerts are sent as JSON POST to this URL. Compatible with Slack, Feishu, Discord, or any webhook receiver.

Quality Tier

quality_tier: high
# high: Opus for consolidation, LLM for ambiguous classification
# medium: Flash for consolidation, keywords-only classification
# low: All rules-based, zero LLM calls

Versioning

schema_version: 1
config_version: 2     # Bumped in v2.0.0